1 //===- LoopLoadElimination.cpp - Loop Load Elimination Pass ---------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implement a loop-aware load elimination pass. 10 // 11 // It uses LoopAccessAnalysis to identify loop-carried dependences with a 12 // distance of one between stores and loads. These form the candidates for the 13 // transformation. The source value of each store then propagated to the user 14 // of the corresponding load. This makes the load dead. 15 // 16 // The pass can also version the loop and add memchecks in order to prove that 17 // may-aliasing stores can't change the value in memory before it's read by the 18 // load. 19 // 20 //===----------------------------------------------------------------------===// 21 22 #include "llvm/Transforms/Scalar/LoopLoadElimination.h" 23 #include "llvm/ADT/APInt.h" 24 #include "llvm/ADT/DenseMap.h" 25 #include "llvm/ADT/DepthFirstIterator.h" 26 #include "llvm/ADT/STLExtras.h" 27 #include "llvm/ADT/SmallPtrSet.h" 28 #include "llvm/ADT/SmallVector.h" 29 #include "llvm/ADT/Statistic.h" 30 #include "llvm/Analysis/AssumptionCache.h" 31 #include "llvm/Analysis/BlockFrequencyInfo.h" 32 #include "llvm/Analysis/GlobalsModRef.h" 33 #include "llvm/Analysis/LazyBlockFrequencyInfo.h" 34 #include "llvm/Analysis/LoopAccessAnalysis.h" 35 #include "llvm/Analysis/LoopAnalysisManager.h" 36 #include "llvm/Analysis/LoopInfo.h" 37 #include "llvm/Analysis/ProfileSummaryInfo.h" 38 #include "llvm/Analysis/ScalarEvolution.h" 39 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 40 #include "llvm/Analysis/TargetLibraryInfo.h" 41 #include "llvm/Analysis/TargetTransformInfo.h" 42 #include "llvm/IR/DataLayout.h" 43 #include "llvm/IR/Dominators.h" 44 #include "llvm/IR/Instructions.h" 45 #include "llvm/IR/Module.h" 46 #include "llvm/IR/PassManager.h" 47 #include "llvm/IR/Type.h" 48 #include "llvm/IR/Value.h" 49 #include "llvm/InitializePasses.h" 50 #include "llvm/Pass.h" 51 #include "llvm/Support/Casting.h" 52 #include "llvm/Support/CommandLine.h" 53 #include "llvm/Support/Debug.h" 54 #include "llvm/Support/raw_ostream.h" 55 #include "llvm/Transforms/Scalar.h" 56 #include "llvm/Transforms/Utils.h" 57 #include "llvm/Transforms/Utils/LoopSimplify.h" 58 #include "llvm/Transforms/Utils/LoopVersioning.h" 59 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" 60 #include "llvm/Transforms/Utils/SizeOpts.h" 61 #include <algorithm> 62 #include <cassert> 63 #include <forward_list> 64 #include <set> 65 #include <tuple> 66 #include <utility> 67 68 using namespace llvm; 69 70 #define LLE_OPTION "loop-load-elim" 71 #define DEBUG_TYPE LLE_OPTION 72 73 static cl::opt<unsigned> CheckPerElim( 74 "runtime-check-per-loop-load-elim", cl::Hidden, 75 cl::desc("Max number of memchecks allowed per eliminated load on average"), 76 cl::init(1)); 77 78 static cl::opt<unsigned> LoadElimSCEVCheckThreshold( 79 "loop-load-elimination-scev-check-threshold", cl::init(8), cl::Hidden, 80 cl::desc("The maximum number of SCEV checks allowed for Loop " 81 "Load Elimination")); 82 83 STATISTIC(NumLoopLoadEliminted, "Number of loads eliminated by LLE"); 84 85 namespace { 86 87 /// Represent a store-to-forwarding candidate. 88 struct StoreToLoadForwardingCandidate { 89 LoadInst *Load; 90 StoreInst *Store; 91 92 StoreToLoadForwardingCandidate(LoadInst *Load, StoreInst *Store) 93 : Load(Load), Store(Store) {} 94 95 /// Return true if the dependence from the store to the load has a 96 /// distance of one. E.g. A[i+1] = A[i] 97 bool isDependenceDistanceOfOne(PredicatedScalarEvolution &PSE, 98 Loop *L) const { 99 Value *LoadPtr = Load->getPointerOperand(); 100 Value *StorePtr = Store->getPointerOperand(); 101 Type *LoadType = getLoadStoreType(Load); 102 103 assert(LoadPtr->getType()->getPointerAddressSpace() == 104 StorePtr->getType()->getPointerAddressSpace() && 105 LoadType == getLoadStoreType(Store) && 106 "Should be a known dependence"); 107 108 // Currently we only support accesses with unit stride. FIXME: we should be 109 // able to handle non unit stirde as well as long as the stride is equal to 110 // the dependence distance. 111 if (getPtrStride(PSE, LoadType, LoadPtr, L) != 1 || 112 getPtrStride(PSE, LoadType, StorePtr, L) != 1) 113 return false; 114 115 auto &DL = Load->getParent()->getModule()->getDataLayout(); 116 unsigned TypeByteSize = DL.getTypeAllocSize(const_cast<Type *>(LoadType)); 117 118 auto *LoadPtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(LoadPtr)); 119 auto *StorePtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(StorePtr)); 120 121 // We don't need to check non-wrapping here because forward/backward 122 // dependence wouldn't be valid if these weren't monotonic accesses. 123 auto *Dist = cast<SCEVConstant>( 124 PSE.getSE()->getMinusSCEV(StorePtrSCEV, LoadPtrSCEV)); 125 const APInt &Val = Dist->getAPInt(); 126 return Val == TypeByteSize; 127 } 128 129 Value *getLoadPtr() const { return Load->getPointerOperand(); } 130 131 #ifndef NDEBUG 132 friend raw_ostream &operator<<(raw_ostream &OS, 133 const StoreToLoadForwardingCandidate &Cand) { 134 OS << *Cand.Store << " -->\n"; 135 OS.indent(2) << *Cand.Load << "\n"; 136 return OS; 137 } 138 #endif 139 }; 140 141 } // end anonymous namespace 142 143 /// Check if the store dominates all latches, so as long as there is no 144 /// intervening store this value will be loaded in the next iteration. 145 static bool doesStoreDominatesAllLatches(BasicBlock *StoreBlock, Loop *L, 146 DominatorTree *DT) { 147 SmallVector<BasicBlock *, 8> Latches; 148 L->getLoopLatches(Latches); 149 return llvm::all_of(Latches, [&](const BasicBlock *Latch) { 150 return DT->dominates(StoreBlock, Latch); 151 }); 152 } 153 154 /// Return true if the load is not executed on all paths in the loop. 155 static bool isLoadConditional(LoadInst *Load, Loop *L) { 156 return Load->getParent() != L->getHeader(); 157 } 158 159 namespace { 160 161 /// The per-loop class that does most of the work. 162 class LoadEliminationForLoop { 163 public: 164 LoadEliminationForLoop(Loop *L, LoopInfo *LI, const LoopAccessInfo &LAI, 165 DominatorTree *DT, BlockFrequencyInfo *BFI, 166 ProfileSummaryInfo* PSI) 167 : L(L), LI(LI), LAI(LAI), DT(DT), BFI(BFI), PSI(PSI), PSE(LAI.getPSE()) {} 168 169 /// Look through the loop-carried and loop-independent dependences in 170 /// this loop and find store->load dependences. 171 /// 172 /// Note that no candidate is returned if LAA has failed to analyze the loop 173 /// (e.g. if it's not bottom-tested, contains volatile memops, etc.) 174 std::forward_list<StoreToLoadForwardingCandidate> 175 findStoreToLoadDependences(const LoopAccessInfo &LAI) { 176 std::forward_list<StoreToLoadForwardingCandidate> Candidates; 177 178 const auto *Deps = LAI.getDepChecker().getDependences(); 179 if (!Deps) 180 return Candidates; 181 182 // Find store->load dependences (consequently true dep). Both lexically 183 // forward and backward dependences qualify. Disqualify loads that have 184 // other unknown dependences. 185 186 SmallPtrSet<Instruction *, 4> LoadsWithUnknownDepedence; 187 188 for (const auto &Dep : *Deps) { 189 Instruction *Source = Dep.getSource(LAI); 190 Instruction *Destination = Dep.getDestination(LAI); 191 192 if (Dep.Type == MemoryDepChecker::Dependence::Unknown) { 193 if (isa<LoadInst>(Source)) 194 LoadsWithUnknownDepedence.insert(Source); 195 if (isa<LoadInst>(Destination)) 196 LoadsWithUnknownDepedence.insert(Destination); 197 continue; 198 } 199 200 if (Dep.isBackward()) 201 // Note that the designations source and destination follow the program 202 // order, i.e. source is always first. (The direction is given by the 203 // DepType.) 204 std::swap(Source, Destination); 205 else 206 assert(Dep.isForward() && "Needs to be a forward dependence"); 207 208 auto *Store = dyn_cast<StoreInst>(Source); 209 if (!Store) 210 continue; 211 auto *Load = dyn_cast<LoadInst>(Destination); 212 if (!Load) 213 continue; 214 215 // Only progagate the value if they are of the same type. 216 if (Store->getPointerOperandType() != Load->getPointerOperandType()) 217 continue; 218 219 Candidates.emplace_front(Load, Store); 220 } 221 222 if (!LoadsWithUnknownDepedence.empty()) 223 Candidates.remove_if([&](const StoreToLoadForwardingCandidate &C) { 224 return LoadsWithUnknownDepedence.count(C.Load); 225 }); 226 227 return Candidates; 228 } 229 230 /// Return the index of the instruction according to program order. 231 unsigned getInstrIndex(Instruction *Inst) { 232 auto I = InstOrder.find(Inst); 233 assert(I != InstOrder.end() && "No index for instruction"); 234 return I->second; 235 } 236 237 /// If a load has multiple candidates associated (i.e. different 238 /// stores), it means that it could be forwarding from multiple stores 239 /// depending on control flow. Remove these candidates. 240 /// 241 /// Here, we rely on LAA to include the relevant loop-independent dependences. 242 /// LAA is known to omit these in the very simple case when the read and the 243 /// write within an alias set always takes place using the *same* pointer. 244 /// 245 /// However, we know that this is not the case here, i.e. we can rely on LAA 246 /// to provide us with loop-independent dependences for the cases we're 247 /// interested. Consider the case for example where a loop-independent 248 /// dependece S1->S2 invalidates the forwarding S3->S2. 249 /// 250 /// A[i] = ... (S1) 251 /// ... = A[i] (S2) 252 /// A[i+1] = ... (S3) 253 /// 254 /// LAA will perform dependence analysis here because there are two 255 /// *different* pointers involved in the same alias set (&A[i] and &A[i+1]). 256 void removeDependencesFromMultipleStores( 257 std::forward_list<StoreToLoadForwardingCandidate> &Candidates) { 258 // If Store is nullptr it means that we have multiple stores forwarding to 259 // this store. 260 using LoadToSingleCandT = 261 DenseMap<LoadInst *, const StoreToLoadForwardingCandidate *>; 262 LoadToSingleCandT LoadToSingleCand; 263 264 for (const auto &Cand : Candidates) { 265 bool NewElt; 266 LoadToSingleCandT::iterator Iter; 267 268 std::tie(Iter, NewElt) = 269 LoadToSingleCand.insert(std::make_pair(Cand.Load, &Cand)); 270 if (!NewElt) { 271 const StoreToLoadForwardingCandidate *&OtherCand = Iter->second; 272 // Already multiple stores forward to this load. 273 if (OtherCand == nullptr) 274 continue; 275 276 // Handle the very basic case when the two stores are in the same block 277 // so deciding which one forwards is easy. The later one forwards as 278 // long as they both have a dependence distance of one to the load. 279 if (Cand.Store->getParent() == OtherCand->Store->getParent() && 280 Cand.isDependenceDistanceOfOne(PSE, L) && 281 OtherCand->isDependenceDistanceOfOne(PSE, L)) { 282 // They are in the same block, the later one will forward to the load. 283 if (getInstrIndex(OtherCand->Store) < getInstrIndex(Cand.Store)) 284 OtherCand = &Cand; 285 } else 286 OtherCand = nullptr; 287 } 288 } 289 290 Candidates.remove_if([&](const StoreToLoadForwardingCandidate &Cand) { 291 if (LoadToSingleCand[Cand.Load] != &Cand) { 292 LLVM_DEBUG( 293 dbgs() << "Removing from candidates: \n" 294 << Cand 295 << " The load may have multiple stores forwarding to " 296 << "it\n"); 297 return true; 298 } 299 return false; 300 }); 301 } 302 303 /// Given two pointers operations by their RuntimePointerChecking 304 /// indices, return true if they require an alias check. 305 /// 306 /// We need a check if one is a pointer for a candidate load and the other is 307 /// a pointer for a possibly intervening store. 308 bool needsChecking(unsigned PtrIdx1, unsigned PtrIdx2, 309 const SmallPtrSetImpl<Value *> &PtrsWrittenOnFwdingPath, 310 const SmallPtrSetImpl<Value *> &CandLoadPtrs) { 311 Value *Ptr1 = 312 LAI.getRuntimePointerChecking()->getPointerInfo(PtrIdx1).PointerValue; 313 Value *Ptr2 = 314 LAI.getRuntimePointerChecking()->getPointerInfo(PtrIdx2).PointerValue; 315 return ((PtrsWrittenOnFwdingPath.count(Ptr1) && CandLoadPtrs.count(Ptr2)) || 316 (PtrsWrittenOnFwdingPath.count(Ptr2) && CandLoadPtrs.count(Ptr1))); 317 } 318 319 /// Return pointers that are possibly written to on the path from a 320 /// forwarding store to a load. 321 /// 322 /// These pointers need to be alias-checked against the forwarding candidates. 323 SmallPtrSet<Value *, 4> findPointersWrittenOnForwardingPath( 324 const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) { 325 // From FirstStore to LastLoad neither of the elimination candidate loads 326 // should overlap with any of the stores. 327 // 328 // E.g.: 329 // 330 // st1 C[i] 331 // ld1 B[i] <-------, 332 // ld0 A[i] <----, | * LastLoad 333 // ... | | 334 // st2 E[i] | | 335 // st3 B[i+1] -- | -' * FirstStore 336 // st0 A[i+1] ---' 337 // st4 D[i] 338 // 339 // st0 forwards to ld0 if the accesses in st4 and st1 don't overlap with 340 // ld0. 341 342 LoadInst *LastLoad = 343 std::max_element(Candidates.begin(), Candidates.end(), 344 [&](const StoreToLoadForwardingCandidate &A, 345 const StoreToLoadForwardingCandidate &B) { 346 return getInstrIndex(A.Load) < getInstrIndex(B.Load); 347 }) 348 ->Load; 349 StoreInst *FirstStore = 350 std::min_element(Candidates.begin(), Candidates.end(), 351 [&](const StoreToLoadForwardingCandidate &A, 352 const StoreToLoadForwardingCandidate &B) { 353 return getInstrIndex(A.Store) < 354 getInstrIndex(B.Store); 355 }) 356 ->Store; 357 358 // We're looking for stores after the first forwarding store until the end 359 // of the loop, then from the beginning of the loop until the last 360 // forwarded-to load. Collect the pointer for the stores. 361 SmallPtrSet<Value *, 4> PtrsWrittenOnFwdingPath; 362 363 auto InsertStorePtr = [&](Instruction *I) { 364 if (auto *S = dyn_cast<StoreInst>(I)) 365 PtrsWrittenOnFwdingPath.insert(S->getPointerOperand()); 366 }; 367 const auto &MemInstrs = LAI.getDepChecker().getMemoryInstructions(); 368 std::for_each(MemInstrs.begin() + getInstrIndex(FirstStore) + 1, 369 MemInstrs.end(), InsertStorePtr); 370 std::for_each(MemInstrs.begin(), &MemInstrs[getInstrIndex(LastLoad)], 371 InsertStorePtr); 372 373 return PtrsWrittenOnFwdingPath; 374 } 375 376 /// Determine the pointer alias checks to prove that there are no 377 /// intervening stores. 378 SmallVector<RuntimePointerCheck, 4> collectMemchecks( 379 const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) { 380 381 SmallPtrSet<Value *, 4> PtrsWrittenOnFwdingPath = 382 findPointersWrittenOnForwardingPath(Candidates); 383 384 // Collect the pointers of the candidate loads. 385 SmallPtrSet<Value *, 4> CandLoadPtrs; 386 for (const auto &Candidate : Candidates) 387 CandLoadPtrs.insert(Candidate.getLoadPtr()); 388 389 const auto &AllChecks = LAI.getRuntimePointerChecking()->getChecks(); 390 SmallVector<RuntimePointerCheck, 4> Checks; 391 392 copy_if(AllChecks, std::back_inserter(Checks), 393 [&](const RuntimePointerCheck &Check) { 394 for (auto PtrIdx1 : Check.first->Members) 395 for (auto PtrIdx2 : Check.second->Members) 396 if (needsChecking(PtrIdx1, PtrIdx2, PtrsWrittenOnFwdingPath, 397 CandLoadPtrs)) 398 return true; 399 return false; 400 }); 401 402 LLVM_DEBUG(dbgs() << "\nPointer Checks (count: " << Checks.size() 403 << "):\n"); 404 LLVM_DEBUG(LAI.getRuntimePointerChecking()->printChecks(dbgs(), Checks)); 405 406 return Checks; 407 } 408 409 /// Perform the transformation for a candidate. 410 void 411 propagateStoredValueToLoadUsers(const StoreToLoadForwardingCandidate &Cand, 412 SCEVExpander &SEE) { 413 // loop: 414 // %x = load %gep_i 415 // = ... %x 416 // store %y, %gep_i_plus_1 417 // 418 // => 419 // 420 // ph: 421 // %x.initial = load %gep_0 422 // loop: 423 // %x.storeforward = phi [%x.initial, %ph] [%y, %loop] 424 // %x = load %gep_i <---- now dead 425 // = ... %x.storeforward 426 // store %y, %gep_i_plus_1 427 428 Value *Ptr = Cand.Load->getPointerOperand(); 429 auto *PtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(Ptr)); 430 auto *PH = L->getLoopPreheader(); 431 assert(PH && "Preheader should exist!"); 432 Value *InitialPtr = SEE.expandCodeFor(PtrSCEV->getStart(), Ptr->getType(), 433 PH->getTerminator()); 434 Value *Initial = new LoadInst( 435 Cand.Load->getType(), InitialPtr, "load_initial", 436 /* isVolatile */ false, Cand.Load->getAlign(), PH->getTerminator()); 437 438 PHINode *PHI = PHINode::Create(Initial->getType(), 2, "store_forwarded", 439 &L->getHeader()->front()); 440 PHI->addIncoming(Initial, PH); 441 PHI->addIncoming(Cand.Store->getOperand(0), L->getLoopLatch()); 442 443 Cand.Load->replaceAllUsesWith(PHI); 444 } 445 446 /// Top-level driver for each loop: find store->load forwarding 447 /// candidates, add run-time checks and perform transformation. 448 bool processLoop() { 449 LLVM_DEBUG(dbgs() << "\nIn \"" << L->getHeader()->getParent()->getName() 450 << "\" checking " << *L << "\n"); 451 452 // Look for store-to-load forwarding cases across the 453 // backedge. E.g.: 454 // 455 // loop: 456 // %x = load %gep_i 457 // = ... %x 458 // store %y, %gep_i_plus_1 459 // 460 // => 461 // 462 // ph: 463 // %x.initial = load %gep_0 464 // loop: 465 // %x.storeforward = phi [%x.initial, %ph] [%y, %loop] 466 // %x = load %gep_i <---- now dead 467 // = ... %x.storeforward 468 // store %y, %gep_i_plus_1 469 470 // First start with store->load dependences. 471 auto StoreToLoadDependences = findStoreToLoadDependences(LAI); 472 if (StoreToLoadDependences.empty()) 473 return false; 474 475 // Generate an index for each load and store according to the original 476 // program order. This will be used later. 477 InstOrder = LAI.getDepChecker().generateInstructionOrderMap(); 478 479 // To keep things simple for now, remove those where the load is potentially 480 // fed by multiple stores. 481 removeDependencesFromMultipleStores(StoreToLoadDependences); 482 if (StoreToLoadDependences.empty()) 483 return false; 484 485 // Filter the candidates further. 486 SmallVector<StoreToLoadForwardingCandidate, 4> Candidates; 487 for (const StoreToLoadForwardingCandidate &Cand : StoreToLoadDependences) { 488 LLVM_DEBUG(dbgs() << "Candidate " << Cand); 489 490 // Make sure that the stored values is available everywhere in the loop in 491 // the next iteration. 492 if (!doesStoreDominatesAllLatches(Cand.Store->getParent(), L, DT)) 493 continue; 494 495 // If the load is conditional we can't hoist its 0-iteration instance to 496 // the preheader because that would make it unconditional. Thus we would 497 // access a memory location that the original loop did not access. 498 if (isLoadConditional(Cand.Load, L)) 499 continue; 500 501 // Check whether the SCEV difference is the same as the induction step, 502 // thus we load the value in the next iteration. 503 if (!Cand.isDependenceDistanceOfOne(PSE, L)) 504 continue; 505 506 assert(isa<SCEVAddRecExpr>(PSE.getSCEV(Cand.Load->getPointerOperand())) && 507 "Loading from something other than indvar?"); 508 assert( 509 isa<SCEVAddRecExpr>(PSE.getSCEV(Cand.Store->getPointerOperand())) && 510 "Storing to something other than indvar?"); 511 512 Candidates.push_back(Cand); 513 LLVM_DEBUG( 514 dbgs() 515 << Candidates.size() 516 << ". Valid store-to-load forwarding across the loop backedge\n"); 517 } 518 if (Candidates.empty()) 519 return false; 520 521 // Check intervening may-alias stores. These need runtime checks for alias 522 // disambiguation. 523 SmallVector<RuntimePointerCheck, 4> Checks = collectMemchecks(Candidates); 524 525 // Too many checks are likely to outweigh the benefits of forwarding. 526 if (Checks.size() > Candidates.size() * CheckPerElim) { 527 LLVM_DEBUG(dbgs() << "Too many run-time checks needed.\n"); 528 return false; 529 } 530 531 if (LAI.getPSE().getUnionPredicate().getComplexity() > 532 LoadElimSCEVCheckThreshold) { 533 LLVM_DEBUG(dbgs() << "Too many SCEV run-time checks needed.\n"); 534 return false; 535 } 536 537 if (!L->isLoopSimplifyForm()) { 538 LLVM_DEBUG(dbgs() << "Loop is not is loop-simplify form"); 539 return false; 540 } 541 542 if (!Checks.empty() || !LAI.getPSE().getUnionPredicate().isAlwaysTrue()) { 543 if (LAI.hasConvergentOp()) { 544 LLVM_DEBUG(dbgs() << "Versioning is needed but not allowed with " 545 "convergent calls\n"); 546 return false; 547 } 548 549 auto *HeaderBB = L->getHeader(); 550 auto *F = HeaderBB->getParent(); 551 bool OptForSize = F->hasOptSize() || 552 llvm::shouldOptimizeForSize(HeaderBB, PSI, BFI, 553 PGSOQueryType::IRPass); 554 if (OptForSize) { 555 LLVM_DEBUG( 556 dbgs() << "Versioning is needed but not allowed when optimizing " 557 "for size.\n"); 558 return false; 559 } 560 561 // Point of no-return, start the transformation. First, version the loop 562 // if necessary. 563 564 LoopVersioning LV(LAI, Checks, L, LI, DT, PSE.getSE()); 565 LV.versionLoop(); 566 567 // After versioning, some of the candidates' pointers could stop being 568 // SCEVAddRecs. We need to filter them out. 569 auto NoLongerGoodCandidate = [this]( 570 const StoreToLoadForwardingCandidate &Cand) { 571 return !isa<SCEVAddRecExpr>( 572 PSE.getSCEV(Cand.Load->getPointerOperand())) || 573 !isa<SCEVAddRecExpr>( 574 PSE.getSCEV(Cand.Store->getPointerOperand())); 575 }; 576 llvm::erase_if(Candidates, NoLongerGoodCandidate); 577 } 578 579 // Next, propagate the value stored by the store to the users of the load. 580 // Also for the first iteration, generate the initial value of the load. 581 SCEVExpander SEE(*PSE.getSE(), L->getHeader()->getModule()->getDataLayout(), 582 "storeforward"); 583 for (const auto &Cand : Candidates) 584 propagateStoredValueToLoadUsers(Cand, SEE); 585 NumLoopLoadEliminted += Candidates.size(); 586 587 return true; 588 } 589 590 private: 591 Loop *L; 592 593 /// Maps the load/store instructions to their index according to 594 /// program order. 595 DenseMap<Instruction *, unsigned> InstOrder; 596 597 // Analyses used. 598 LoopInfo *LI; 599 const LoopAccessInfo &LAI; 600 DominatorTree *DT; 601 BlockFrequencyInfo *BFI; 602 ProfileSummaryInfo *PSI; 603 PredicatedScalarEvolution PSE; 604 }; 605 606 } // end anonymous namespace 607 608 static bool 609 eliminateLoadsAcrossLoops(Function &F, LoopInfo &LI, DominatorTree &DT, 610 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, 611 ScalarEvolution *SE, AssumptionCache *AC, 612 function_ref<const LoopAccessInfo &(Loop &)> GetLAI) { 613 // Build up a worklist of inner-loops to transform to avoid iterator 614 // invalidation. 615 // FIXME: This logic comes from other passes that actually change the loop 616 // nest structure. It isn't clear this is necessary (or useful) for a pass 617 // which merely optimizes the use of loads in a loop. 618 SmallVector<Loop *, 8> Worklist; 619 620 bool Changed = false; 621 622 for (Loop *TopLevelLoop : LI) 623 for (Loop *L : depth_first(TopLevelLoop)) { 624 Changed |= simplifyLoop(L, &DT, &LI, SE, AC, /*MSSAU*/ nullptr, false); 625 // We only handle inner-most loops. 626 if (L->isInnermost()) 627 Worklist.push_back(L); 628 } 629 630 // Now walk the identified inner loops. 631 for (Loop *L : Worklist) { 632 // Match historical behavior 633 if (!L->isRotatedForm() || !L->getExitingBlock()) 634 continue; 635 // The actual work is performed by LoadEliminationForLoop. 636 LoadEliminationForLoop LEL(L, &LI, GetLAI(*L), &DT, BFI, PSI); 637 Changed |= LEL.processLoop(); 638 } 639 return Changed; 640 } 641 642 namespace { 643 644 /// The pass. Most of the work is delegated to the per-loop 645 /// LoadEliminationForLoop class. 646 class LoopLoadElimination : public FunctionPass { 647 public: 648 static char ID; 649 650 LoopLoadElimination() : FunctionPass(ID) { 651 initializeLoopLoadEliminationPass(*PassRegistry::getPassRegistry()); 652 } 653 654 bool runOnFunction(Function &F) override { 655 if (skipFunction(F)) 656 return false; 657 658 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 659 auto &LAA = getAnalysis<LoopAccessLegacyAnalysis>(); 660 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 661 auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); 662 auto *BFI = (PSI && PSI->hasProfileSummary()) ? 663 &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI() : 664 nullptr; 665 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 666 667 // Process each loop nest in the function. 668 return eliminateLoadsAcrossLoops( 669 F, LI, DT, BFI, PSI, SE, /*AC*/ nullptr, 670 [&LAA](Loop &L) -> const LoopAccessInfo & { return LAA.getInfo(&L); }); 671 } 672 673 void getAnalysisUsage(AnalysisUsage &AU) const override { 674 AU.addRequiredID(LoopSimplifyID); 675 AU.addRequired<LoopInfoWrapperPass>(); 676 AU.addPreserved<LoopInfoWrapperPass>(); 677 AU.addRequired<LoopAccessLegacyAnalysis>(); 678 AU.addRequired<ScalarEvolutionWrapperPass>(); 679 AU.addRequired<DominatorTreeWrapperPass>(); 680 AU.addPreserved<DominatorTreeWrapperPass>(); 681 AU.addPreserved<GlobalsAAWrapperPass>(); 682 AU.addRequired<ProfileSummaryInfoWrapperPass>(); 683 LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU); 684 } 685 }; 686 687 } // end anonymous namespace 688 689 char LoopLoadElimination::ID; 690 691 static const char LLE_name[] = "Loop Load Elimination"; 692 693 INITIALIZE_PASS_BEGIN(LoopLoadElimination, LLE_OPTION, LLE_name, false, false) 694 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 695 INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis) 696 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 697 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 698 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 699 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass) 700 INITIALIZE_PASS_DEPENDENCY(LazyBlockFrequencyInfoPass) 701 INITIALIZE_PASS_END(LoopLoadElimination, LLE_OPTION, LLE_name, false, false) 702 703 FunctionPass *llvm::createLoopLoadEliminationPass() { 704 return new LoopLoadElimination(); 705 } 706 707 PreservedAnalyses LoopLoadEliminationPass::run(Function &F, 708 FunctionAnalysisManager &AM) { 709 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); 710 auto &LI = AM.getResult<LoopAnalysis>(F); 711 auto &TTI = AM.getResult<TargetIRAnalysis>(F); 712 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 713 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 714 auto &AA = AM.getResult<AAManager>(F); 715 auto &AC = AM.getResult<AssumptionAnalysis>(F); 716 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F); 717 auto *PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent()); 718 auto *BFI = (PSI && PSI->hasProfileSummary()) ? 719 &AM.getResult<BlockFrequencyAnalysis>(F) : nullptr; 720 721 auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager(); 722 bool Changed = eliminateLoadsAcrossLoops( 723 F, LI, DT, BFI, PSI, &SE, &AC, [&](Loop &L) -> const LoopAccessInfo & { 724 LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, 725 TLI, TTI, nullptr, nullptr, nullptr}; 726 return LAM.getResult<LoopAccessAnalysis>(L, AR); 727 }); 728 729 if (!Changed) 730 return PreservedAnalyses::all(); 731 732 PreservedAnalyses PA; 733 return PA; 734 } 735